Voice keyword recognition method and system based on multi-scale global convolutional network

Abstract

The invention discloses a voice keyword recognition method and system based on a multi-scale global convolutional network, and relates to the technical field of voice recognition, and the method comprises the following steps: carrying out the preprocessing of voice information through a pre-convolution block, and obtaining a preliminary voice feature; in the multi-scale feature fusion residual module, the preliminary voice features are processed, and interference of irrelevant information on the voice features is reduced at the same time; in the final convolution block, the speech features output by the multi-scale feature fusion residual module are processed to obtain final speech features; and obtaining a keyword recognition result after the final speech features pass through a maximum pooling layer and a full connection layer. According to the invention, deeper voice information is extracted in the time domain and the frequency domain, so that mutual interference possibly existing between information in different fields is avoided; time domain global convolution and frequency domain global convolution can capture global features of voice information in a time domain and a frequency domain respectively, and the global feature perception ability of the model is enhanced to learn richer feature representation.

Description

Technical Field

[0001] The present invention relates to the field of speech recognition technology, and in particular to a speech keyword recognition method and system based on a multi-scale global convolutional network. Background Art

[0002] Deep learning algorithms have entered a new stage, surpassing human accuracy in cognitive tasks such as speech recognition. Many artificial intelligence products have been integrated into daily life, significantly improving the efficiency of people's work, study and life. Voice interaction technology frees hands and eyes, and converts voice into text to control devices, such as voice assistants, command control, smart homes and mobile phones. It is particularly efficient in scenarios that require multi-sensory collaborative operation. Voice interaction devices are mostly in standby mode, monitoring wake-up words to activate interactive programs. This technology is leading the trend of future technological life, evolving from an auxiliary tool to an independent product, demonstrating its rapid development.

[0003] Most voice interaction products are small embedded devices that use voice keyword recognition models to detect wake-up words. The model must have small memory usage, low computational effort, and high accuracy to save device power. However, there is a huge difference between testing in a quiet laboratory environment and the noise interference in a real environment, which leads to a significant decrease in model accuracy. Therefore, maintaining high performance in a noisy environment is crucial, as it affects whether users can interact with the device smoothly in an open scene. Given that noise is unavoidable in real life, building a lightweight and noise-robust voice keyword recognition model has become an urgent problem to be solved.

[0004] Speech keyword recognition models based on deep learning improve accuracy through deep neural network architecture, data sets, and data enhancement techniques. These models learn speech feature representations and adapt to various pronunciations, accents, and speaking styles through large-scale data set training to improve robustness. Data enhancement techniques introduce noise and transformations to enable the model to perform well in noisy environments. Optimizing model structure and parameters can reduce memory usage and achieve a lighter model. Currently, the technologies are mainly divided into two categories:

[0005] The first is the convolution-based speech keyword recognition model, such as BC-ResNet, which uses convolution kernels to capture local features and frequency information of audio signals, and has the advantages of fewer parameters and high computational efficiency. However, the convolution kernel only senses local information and may ignore the global context, resulting in low robustness and susceptibility to noise, accent or speech speed changes.

[0006] The second is the speech keyword recognition model based on the attention model, such as the AST model, which dynamically allocates weights to focus on important information in the input sequence, captures long-distance context and key speech signals, and adapts to noisy environments and accent changes. However, it has many parameters, occupies a lot of memory, increases the risk of overfitting, and requires more data for effective training.

[0007] In addition, reverberation (sound wave reflection, refraction, and attenuation) in everyday scenes can also reduce recognition accuracy. Reverberation causes the original sound to overlap with the reflected sound, and the overlap usually occurs locally. Noise manifests itself as amplitude fluctuations in the time domain and additional frequency components in the frequency domain. However, the serial stacking frequency domain convolution and time domain convolution method used in the current model may cause information loss and affect subsequent feature extraction. Summary of the invention

[0008] The purpose of the present invention is to propose a speech keyword recognition method and system based on a multi-scale global convolutional network to improve the accuracy of keyword recognition and reduce the number of parameters while ensuring recognition performance.

[0009] According to a first aspect of an embodiment of the present disclosure, a method for speech keyword recognition based on a multi-scale global convolutional network is provided, comprising the following steps:

[0010] Preprocess the speech information through the pre-convolution block to obtain preliminary speech features;

[0011] In the multi-scale feature fusion residual module, the preliminary speech features are processed while reducing the interference of irrelevant information on the speech features;

[0012] In the final convolution block, the speech features output by the multi-scale feature fusion residual module are processed to obtain the final speech features;

[0013] The final speech features are passed through the maximum pooling layer and the fully connected layer to obtain the keyword recognition result.

[0014] According to a second aspect of an embodiment of the present disclosure, a speech keyword recognition system based on a multi-scale global convolutional network is provided, comprising:

[0015] Pre-convolution block, pre-processes the speech information to obtain preliminary speech features;

[0016] Multi-scale feature fusion residual module, which processes preliminary speech features while reducing the interference of irrelevant information on speech features;

[0017] The final convolution block processes the speech features output by the multi-scale feature fusion residual module to obtain the final speech features;

[0018] Recognition module, the final speech features pass through the maximum pooling layer and the fully connected layer to obtain the keyword recognition result.

[0019] According to a third aspect of an embodiment of the present disclosure, an electronic device is provided, comprising a memory, a processor, and a computer program stored and running on the memory, wherein when the processor executes the program, the method for speech keyword recognition based on a multi-scale global convolutional network is implemented.

[0020] According to a fourth aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided, on which a computer program is stored. When the program is executed by a processor, the method for speech keyword recognition based on a multi-scale global convolutional network is implemented.

[0021] Compared with the prior art, the above technical solution adopted by the present invention has the following advantages: the present invention constructs a speech keyword recognition method based on a multi-scale global convolutional network, and proposes parallel multi-scale time domain convolution and multi-scale frequency domain convolution, parallel time domain global convolution and frequency domain global convolution to address the problem of low robustness of keyword recognition under noise conditions. The proposed multi-scale time domain convolution and multi-scale frequency domain convolution separate the extraction process of time domain and frequency domain features in a parallel manner, and use multi-scale convolution to fuse features of different scales to increase feature richness, thereby improving the accuracy of keyword recognition under different noise conditions. At the same time, the proposed time domain global convolution and frequency domain global convolution regard speech frames and frequency bands as the basic units of feature extraction, enhance the model's perception of global features, and further improve the recognition performance under low signal-to-noise ratio. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The drawings in the specification, which constitute a part of the present application, are used to provide further understanding of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute improper limitations on the present application.

[0023] Figure 1 This is a flow chart of a speech keyword recognition method based on a multi-scale global convolutional network;

[0024] Figure 2 Schematic diagram of the multi-scale global convolutional network structure;

[0025] Figure 3 Schematic diagram of multi-scale feature fusion residual module;

[0026] Figure 4 Schematic diagram of time domain global convolution and frequency domain global convolution modules. DETAILED DESCRIPTION

[0027] The present disclosure is further described below in conjunction with the accompanying drawings and embodiments.

[0028] It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0029] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0030] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the possible implementation architecture, functions and operations of the methods and systems according to various embodiments of the present disclosure. It should be noted that each box in the flowchart or block diagram can represent a module, a program segment, or a part of a code, and the module, program segment, or a part of a code may include one or more executable instructions for implementing the logical functions specified in each embodiment. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, or they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the flowchart and / or block diagram, and the combination of boxes in the flowchart and / or block diagram can be implemented using a dedicated hardware-based system that performs a specified function or operation, or can be implemented using a combination of dedicated hardware and computer instructions.

[0031] Embodiment 1:

[0032] like Figure 1 As shown in the figure, first the original audio is preprocessed, including pre-emphasis of the original speech signal, framing of the pre-emphasized speech signal, and windowing of each frame. Then each frame of the speech signal is transformed by short-time Fourier transform to obtain a spectrogram. After that, the spectrogram passes through a 64-dimensional Mel filter bank, and the logarithm along the frequency dimension is taken to obtain the FBank feature. The FBank feature is input into MGConvNet for feature extraction. The features extracted by the model are subjected to maximum pooling, fully connected layers, and Softmax, and finally the classification probability of the specified label is output. The maximum value of the probability represents the predicted label of the current speech.

[0033] like Figure 2 As shown, this embodiment provides a speech keyword recognition method based on a multi-scale global convolutional network, comprising the following steps:

[0034] S1. Preprocess the speech information through the pre-convolution block to obtain preliminary speech features;

[0035] Specifically, the pre-convolution block includes multiple convolutional layers, each of which includes two-dimensional convolution, batch normalization, and Swish activation function. Four convolutional layers are used as an example to illustrate parameter settings:

[0036] First layer: convolution kernel = 3, number of channels = 1, stride = 1;

[0037] Second layer: convolution kernel = 3, number of channels = 32, stride = 1;

[0038] The third layer: convolution kernel = 3, number of channels = 32, stride = 1;

[0039] Fourth layer: convolution kernel = 1, number of channels = 1, stride = 1.

[0040] S2. In the multi-scale feature fusion residual module, the preliminary speech features are processed while reducing the interference of irrelevant information on the speech features;

[0041] Specifically, the multi-scale feature fusion residual module is set in multiple, such as Figure 3 As shown in Figure 1, each multi-scale feature fusion residual module includes multi-scale time domain convolution, multi-scale frequency domain convolution, time domain global convolution, and frequency domain global convolution. The preliminary speech features first undergo multi-scale time domain convolution and multi-scale frequency domain convolution, and then enter the time domain global convolution and frequency domain global convolution. The specific structure is as follows: Figure 4 As shown in Table 1, the output features are obtained after splicing and fusion. The parameter settings are described by taking four multi-scale feature fusion residual modules as an example, as shown in Table 1.

[0042] Table 1 Multi-scale parallel convolution parameters

[0043]

[0044] The overall process of implementing the multi-scale feature fusion residual module is as follows:

[0045]

[0046] Z t' =δ(LN(Z t +BC(C time (Z t )))) (5)

[0047] Z f' =δ(LN(Z f +BC((C freq ((Z f ) T )) T ))) (6)

[0048]

[0049] Among them, formula (1)(2)(3) shows the calculation process of multi-scale time domain convolution. In order to ensure that the feature shape after convolution remains unchanged, the negative values ​​in the formula are filled with 0, and the filling size used is a = (d-1) / 2. Formula (1) shows the calculation process of channel-by-channel convolution, I k,j+i represents the (j+i)th element in the kth channel of input feature I, represents the (i+p+1)th element in the kth filter, δ represents the Swish activation function, represents the jth element in the kth channel of the output feature. Formula (2) shows the calculation process of point-by-point convolution, where represents the first element of the i-th channel in the k-th filter, δ represents the Swish activation function, The jth element in the kth channel of the output feature also represents the final output of the feature after time-domain convolution. Formula (3) fuses the outputs of two time-domain convolutions of different scales, where Represents the element in the i-th row and j-th column of the output feature, Represents the value of the i-th row and j-th column of the k-th channel in the feature after channel connection. Represents the first value in the kth channel in the 2D filter, and BN stands for batch normalization.

[0050] The multi-scale frequency domain convolution process is similar to the multi-scale time domain convolution process. Equation (4) shows the process of fusing the outputs of two frequency domain convolutions of different scales, where Represents the element in the i-th row and j-th column of the output feature, represents the eigenvalue of the i-th row and j-th column in the k-th channel after the channels are connected. Represents the first value of the kth channel in the two-dimensional filter, and BN represents batch normalization.

[0051] Formula (5) represents the calculation process of global convolution in time domain. time In, Z t First enter C time In the example, the output feature shape is The broadcast operation then copies it to match The original shape and Z t Finally, the output Z is obtained through normalization and Swish activation function. t '. BC stands for broadcast operation, the feature expansion will be the specified shape and added; LN stands for layer normalization; δ represents the Swish activation function.

[0052] Formula (6) represents the calculation process of frequency domain global convolution. Frequency domain global convolution C freq In, Z fAt input to C freq The feature will undergo a transposition operation before it is transformed into a global convolution in the frequency domain, extracting all the information of adjacent frequency bands in units of frequency bands. After expansion, it is fused into Z f In the process, the fusion of frequency domain global information is completed. After layer normalization and Swish activation function, the final output feature Z f '.

[0053] Formula (7) represents the point-by-point convolution operation after the feature undergoes global convolution in the time domain and global convolution in the frequency domain, thereby fusing feature information of two different scales. t ' and Z t 'Connect along the channel dimension, expressed as Finally, the feature fusion between time domain and frequency domain is realized through a 2D convolution layer with a convolution kernel size of 1×1. The convolution is expressed as The output features are expressed as Z 1,i,j Represents the element in the i-th row and j-th column of the output feature, represents the eigenvalue of the i-th row and j-th column in the k-th channel after channel connection, and Represents the first value of the kth channel in the convolution filter.

[0054] S3. In the final convolution block, the speech features output by the multi-scale feature fusion residual module are processed to obtain the final speech features;

[0055] Specifically, the final convolution block includes multiple depth-wise separable convolutions.

[0056] S4. After the final speech features pass through the maximum pooling layer and the fully connected layer, the keyword recognition result is obtained.

[0057] According to the noise conditions in daily life and the voice wake-up word triggering scenarios, the following experiments were designed to analyze its beneficial effects:

[0058] The Google Speech CommandV2-12 dataset, MUSAN dataset, and BUT Speech@FITReverberation Database dataset were used for keyword spotting experiments.

[0059] The Google Speech CommandV2-12 (GSCV2) dataset contains 105,000 words and 35 categories in total. Each word is 1 second long and is sampled at 16kHz. Following the official data division, the data division includes training, validation, and test sets. 12 labels were used for classification tasks, including "up", "down", "left", "right", "yes", "no", "on", "off", "go", "stop", as well as "silence" and "unknown" categories. The latter two categories were generated from the remaining words in the dataset, and the number of voices in the "unknown" and "silence" categories was rebalanced to ensure a balanced sample distribution.

[0060] In order to simulate the noise environment in real life more realistically, the MUSAN dataset is used. This dataset contains 930 various noise samples, which are sampled at 16kHz and have a total duration of about 6 hours. The noise samples in the MUSAN dataset cover a variety of technical and non-technical noises, such as DTMF tones, thunder, car horns, etc. These noise samples are added to the speech keywords to simulate audio under different noise conditions.

[0061] BUT Speech@FIT Reverberation Database is a reverberation dataset designed to support research in the field of speech signal processing, especially tasks related to reverberation. It includes reverberation data collected from nine rooms of different sizes, capturing the reflection characteristics of sound propagation in each room. This dataset is of great value to the research of acoustic simulation, audio processing, and audio system performance evaluation.

[0062] In this embodiment, a 64-dimensional Log Mel FilterBank (FBank) feature is used as input, with a window size of 25 milliseconds and an offset of 10 milliseconds. The feature dimension of FBank is fixed to R 98×64, where 98 represents the number of speech frames in each utterance. For keywords with less than 98 speech frames, 0 is used to fill the right side to the appropriate length. In order to increase the diversity of the data, data augmentation was performed during the training process, where the time offset range was -100 milliseconds to 100 milliseconds. In addition, time masking and frequency masking were used, where the maximum length of time masking was 25 milliseconds. All experiments used the same data augmentation settings. In order to enhance the robustness of the model, a multi-condition training method was used. The noise data from MUSAN and the room impulse response data from BUT Speech@FIT Reverberation Database were pre-merged into the training set of GSCV2 to generate speech data with different SNRs and RIRs. The noise data from MUSAN and the RIR data from BUT Speech@FIT Reverberation Database were divided into three parts in a ratio of 8:1:1. These parts were used to mix into the training, validation, and test sets of GSCV2, and the original speech in GSCV2 was marked as "clean" speech, indicating unprocessed pure speech. The final training set contains speech data of five different SNR types: [clean, RIR, 0, -5, -10]dB, with the same number of instances for each type. To ensure the effectiveness of the model for validation experiments, the validation set of GSCV2 also underwent the same processing, including five different SNR types. The difference is that in the test phase, this chapter sets up six separate test phases, each corresponding to generating one type of data from [clean, RIR, 10, 0, -5, -10]dB. In addition, test data with an SNR of 10dB is introduced. This type of data was not involved in the training during the training phase, aiming to evaluate the generalization ability of the model in SNR speech that has not been trained.

[0063] The model is trained with a batch size of 128 and an initial learning rate of 6e -3 Start training. During the training process, the Adam optimizer and the binary cross entropy loss function were used. In order to optimize the training process, an early stopping mechanism was set to prevent the model from overfitting. When the accuracy of the validation set did not increase in ten consecutive rounds, the training was stopped. The model with the highest accuracy of the validation set was saved as the optimal model, and the final test was performed on the test set. In order to reflect the low parameter amount and high model robustness of the model in the present invention, it was compared with the current mainstream keyword recognition model, and the final test results are shown in Table 2.

[0064] Table 2 Model test results under different signal-to-noise ratios

[0065]

[0066] It can be observed from the table that when evaluated under noisy noise conditions, the MGConvNet model proposed in the present invention not only reduces the number of parameters but also improves the accuracy, especially in the case of low signal-to-noise ratio, showing a more obvious effect. This shows that the method of the present invention is more effective in noise robustness and can be more flexibly deployed on small terminal devices.

[0067] Embodiment 2:

[0068] This embodiment provides a speech keyword recognition system based on a multi-scale global convolutional network, including:

[0069] Pre-convolution block, pre-processes the speech information to obtain preliminary speech features;

[0070] Multi-scale feature fusion residual module, which processes preliminary speech features while reducing the interference of irrelevant information on speech features;

[0071] The final convolution block processes the speech features output by the multi-scale feature fusion residual module to obtain the final speech features;

[0072] Recognition module, the final speech features pass through the maximum pooling layer and the fully connected layer to obtain the keyword recognition result.

[0073] Embodiment three:

[0074] An electronic device includes a memory, a processor, and a computer program stored and running on the memory, wherein the processor implements the above-mentioned method for speech keyword recognition based on a multi-scale global convolutional network when executing the program, including:

[0075] Preprocess the speech information through the pre-convolution block to obtain preliminary speech features;

[0076] In the multi-scale feature fusion residual module, the preliminary speech features are processed while reducing the interference of irrelevant information on the speech features;

[0077] In the final convolution block, the speech features output by the multi-scale feature fusion residual module are processed to obtain the final speech features;

[0078] The final speech features are passed through the maximum pooling layer and the fully connected layer to obtain the keyword recognition result.

[0079] Embodiment 4:

[0080] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned method for speech keyword recognition based on a multi-scale global convolutional network, comprising:

[0081] Preprocess the speech information through the pre-convolution block to obtain preliminary speech features;

[0082] In the multi-scale feature fusion residual module, the preliminary speech features are processed while reducing the interference of irrelevant information on the speech features;

[0083] In the final convolution block, the speech features output by the multi-scale feature fusion residual module are processed to obtain the final speech features;

[0084] The final speech features are passed through the maximum pooling layer and the fully connected layer to obtain the keyword recognition result.

[0085] The multi-scale time domain convolution and multi-scale frequency domain convolution in the present invention can extract deeper speech information in the time domain and frequency domain respectively according to the different effects of noise on the time domain and frequency domain, avoid the possible mutual interference between information in different fields, and multi-scale convolution can better integrate features of different scales in the time domain and frequency domain, thereby improving the feature expression ability of the model. The time domain global convolution and the frequency domain global convolution can capture the global features of speech information in the time domain and frequency domain respectively, enhance the global feature perception ability of the model to learn richer feature representations.

[0086] Those skilled in the art should understand that the modules or steps of the present disclosure can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present disclosure is not limited to any specific combination of hardware and software.

[0087] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0088] Although the above describes the specific implementation methods of the present disclosure in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present disclosure. Technical personnel in the relevant field should understand that on the basis of the technical solution of the present disclosure, various modifications or variations that can be made by those skilled in the art without creative work are still within the scope of protection of the present disclosure.

Claims

1. A speech keyword recognition method based on a multi-scale global convolutional network, characterized in that: The following steps are involved: Preprocess the speech information through the pre-convolution block to obtain preliminary speech features; In the multi-scale feature fusion residual module, the preliminary speech features are processed while reducing the interference of irrelevant information on the speech features; In the final convolution block, the speech features output by the multi-scale feature fusion residual module are processed to obtain the final speech features; The final speech features are passed through the maximum pooling layer and the fully connected layer to obtain the keyword recognition result.

2. The method for speech keyword recognition based on a multi-scale global convolutional network according to claim 1, characterized in that: The pre-convolution block includes multiple convolution layers, each of which includes a two-dimensional convolution, a batch normalization, and a Swish activation function.

3. The method for speech keyword recognition based on a multi-scale global convolutional network according to claim 1, characterized in that: The multi-scale feature fusion residual module is provided in plurality.

4. The method for speech keyword recognition based on a multi-scale global convolutional network according to claim 1 or 3, characterized in that: The multi-scale feature fusion residual module includes multi-scale time domain convolution, multi-scale frequency domain convolution, time domain global convolution, and frequency domain global convolution. The preliminary speech features first undergo multi-scale time domain convolution and multi-scale frequency domain convolution, then enter the time domain global convolution and frequency domain global convolution, and finally are spliced ​​and fused to obtain the output features.

5. The method for speech keyword recognition based on a multi-scale global convolutional network according to claim 4, characterized in that: The multi-scale time domain convolution and multi-scale frequency domain convolution representation process is: Among them, formula (1) shows the calculation process of multi-scale time domain convolution; Represents the element in the i-th row and j-th column of the output feature, Represents the value of the i-th row and j-th column of the k-th channel in the feature after channel connection. Represents the first value in the kth channel in the 2D filter, BN represents batch normalization, and δ represents the Swish activation function; Formula (2) shows the calculation process of multi-scale frequency domain convolution; Represents the element in the i-th row and j-th column of the output feature, represents the eigenvalue of the i-th row and j-th column in the k-th channel after the channels are connected. represents the first value of the kth channel in the two-dimensional filter, BN represents batch normalization, and δ represents the Swish activation function.

6. The method for speech keyword recognition based on a multi-scale global convolutional network according to claim 4, characterized in that: The time domain global convolution and frequency domain global convolution representation process are as follows: WITH t' =δ(LN(Z t +BC(C time (WITH t )))) (3) WITH f' =δ(LN(Z f +BC((C freq ((WITH f ) T )) T ))) (4) Among them, formula (3) represents the calculation process of global convolution in time domain. In the global convolution in time domain C time In, Z t First enter C time In the example, the output feature shape is The broadcast operation then copies it to match The original shape and the feature Z t Add to perform feature fusion; finally, obtain the output feature Z through normalization and Swish activation function t '; BC represents broadcast operation; LN represents layer normalization; δ represents Swish activation function; Formula (4) represents the calculation process of frequency domain global convolution. Frequency domain global convolution C freq In, Z f At input to C freq The feature will undergo a transposition operation before it is transformed into a feature Z. After that, the feature undergoes a global convolution in the frequency domain to extract all the information of adjacent frequency bands in units of frequency bands. After expansion, it is fused into feature Z. f In the process, the fusion of global information in the frequency domain is completed; after layer normalization and Swish activation function, the final output feature Z f '.

7. The method for speech keyword recognition based on a multi-scale global convolutional network according to claim 4, characterized in that: After global convolution in time domain and global convolution in frequency domain, point-by-point convolution operation is performed to fuse feature information of two different scales: Among them, Z t ' and Z f 'Connect along the channel dimension, expressed as The feature fusion between time domain and frequency domain is realized through the convolution layer, and the convolution is expressed as The output features are expressed as Z 1,i,j Represents the element in the i-th row and j-th column of the output feature, represents the eigenvalue of the i-th row and j-th column in the k-th channel after channel connection, and Represents the first value of the kth channel in the convolution filter.

8. A speech keyword recognition system based on a multi-scale global convolutional network, characterized in that: include: Pre-convolution block, pre-processes the speech information to obtain preliminary speech features; Multi-scale feature fusion residual module, which processes preliminary speech features while reducing the interference of irrelevant information on speech features; The final convolution block processes the speech features output by the multi-scale feature fusion residual module to obtain the final speech features; Recognition module, the final speech features pass through the maximum pooling layer and the fully connected layer to obtain the keyword recognition result.

9. An electronic device comprising a memory, a processor and a computer program stored and running on the memory, characterized in that: When the processor executes the program, the speech keyword recognition method based on a multi-scale global convolutional network described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the speech keyword recognition method based on a multi-scale global convolutional network described in any one of claims 1 to 7 is implemented.

Images